39
$\begingroup$

It seems "common knowledge" that the following holds:

Let $T$ be a linear transformation on $n\times n$ matrices with complex coefficients that preserves the determinant. Then there exists matrices $U$ and $V$ whose product has determinant 1 such that one of the following holds:

a) For any matrix $A$ we have $T(A)=UAV$
b) For any matrix $A$ we have $T(A)=UA^TV$ where $A^T$ is the transpose of $A$

It seems quite reasonable, but as far as "common knowledge" goes, I have no clue right now on how to prove such a thing?

$\endgroup$
1
  • 6
    $\begingroup$ This is the Frobenius determinant preserver theorem, generalized in a 1949 paper by Dieudonne. It is true over any field. $\endgroup$
    – Ben McKay
    May 8, 2023 at 9:02

6 Answers 6

20
$\begingroup$

First, some easy observations: $T$ must be injective since for any $A$, there is some $B$ such that $B$ and $A+B$ have different determinants (easy exercise). By multiplying $T$ by $T(1)^{-1}$, it may be assumed that $T(1)=1$.

Now note that $T$ preserves the rank of matrices. Indeed, $T$ must preserve the rank $n$ matrices, and then the rank $n-1$ matrices are just the nonsingular locus in the variety of matrices with determinant $0$. This implies $T$ preserves rank $n-1$ matrices. Rank $n-2$ matrices are then the nonsingular locus in rank $<n-1$ matrices so they are preserved, and so on.

Now rank $k$ projections are exactly those rank $k$ matrices which when subtracted from the identity give you something of rank $n-k$; this is easy to see from Jordan normal form. Thus $T$ sends rank $1$ projections to rank $1$ projections. Two projections have disjoint ranges and commute iff their sum is also a projection. In particular, for $P_i$ the projections onto a basis $e_i$, $T$ sends $P_i$ to projections $Q_i$ onto some other basis $f_i$. Now let $U$ be the change of basis matrix from the $e_i$ to the $f_i$. Conjugating $T$ by $U$ shows that we may assume $T$ fixes each $P_i$. That is, picking the standard basis, $T$ fixes all diagonal matrices.

Now matrices whose only nonzero entries are either all in the first row or all in the first column are characterized by the fact that they are rank $1$ and they remain rank $1$ if their first diagonal entry changes. Similar statements hold for other rows and columns. It follows that $T(e_{ij})$ is a multiple of either $e_{ij}$ or $e_{ji}$ for all $j$ and $i$, where $e_{ij}$ is the matrix with $ij$ entry $1$ and all others $0$. By considering the ranks of matrices with only two nonzero entries, it is now easy to see that we must either always have $T(e_{ij})$ a multiple of $e_{ij}$ or always have $T(e_{ij})$ a multiple of $e_{ji}$. Composing $T$ with the transpose map we may assume we are in the first case.

Now let $a_{ij}$ be the scalars such that $T(e_{ij})=a_{ij}e_{ij}$. We know that $a_{ii}=1$, and by considering permutation matrices, it is easy to see that we must have $a_{ij}a_{jk}=a_{ik}$. It follows that $T$ coincides with conjugation by the diagonal matrix with diagonal entries $a_{1i}$, and in particular $T$ has the form $T(A)=UAV$.

$\endgroup$
3
  • 3
    $\begingroup$ Well, just to nitpick a bit, there are some minor wrinkles, like if $T\begin{pmatrix} a & b \\ c & d \end{pmatrix}= \begin{pmatrix} a & -b \\ -c & d \end{pmatrix}$. Then we would have to change $T$ one more time by conjugating with $\operatorname{diag}(1,-1)$ or similar. $\endgroup$ Jan 24, 2020 at 14:48
  • 3
    $\begingroup$ You're right, and apparently it took over 10 years for anyone to catch that! I've corrected the argument at the end. $\endgroup$ Jan 24, 2020 at 15:46
  • $\begingroup$ The answer is positive over any algebraically closed field. See the following question at math.stackexchange: link. $\endgroup$ May 4, 2020 at 12:54
11
$\begingroup$

The conclusion you indicate is obtained as the main result in the following paper, but with an apparently stronger hypothesis: (EDIT: Not stronger at all, actually - just realized you're assuming the map is linear.)

Determinant preserving maps on matrix algebras

Gregor Dolinar and Peter Semrl

Linear Algebra and its Applications Volume 348, Issues 1-3, 15 June 2002, Pages 189-192, DOI: 10.1016/S0024-3795(01)00578-X

Let $M_n$ be the algebra of all $n\times n$ complex matrices. If $\phi:M_n→M_n$ is a surjective mapping satisfying $\det(A+\lambda B)=\det(\phi(A)+\lambda\phi(B))$ then either $\phi$ is of the form $\phi(A)=MAN$ or $\phi$ is of the form $\phi(A)=MA^TN$ where $M,N$ are nonsingular matrices with $\det(MN)=1$.

$\endgroup$
6
  • $\begingroup$ It seems to me that the added hypothesis follows from the requirement that phi is linear. Am I missing something? $\endgroup$ Oct 15, 2009 at 0:11
  • $\begingroup$ Of course - just missed that. It may actually make the claim easier to prove, I'm not sure. $\endgroup$
    – Alon Amit
    Oct 15, 2009 at 1:14
  • $\begingroup$ link does not work $\endgroup$
    – user76479
    Oct 17, 2015 at 21:12
  • 2
    $\begingroup$ @Arul, I updated the link. $\endgroup$
    – Alon Amit
    Oct 17, 2015 at 21:16
  • 5
    $\begingroup$ Unfortunately, this article does not prove the theorem OP is asking about, i.e., Frobenius's linear determinant preserver theorem. In fact the article uses Frobenius's theorem in its proof. The basic point of the article is to show that the seemingly weaker hypothesis actually implies linearity, and then Frobenius's theorem can be applied. $\endgroup$ Jan 24, 2020 at 8:21
8
$\begingroup$

Eric Wofsey's answer can be modified a bit to drop the assumption that we're dealing with the field $K=\mathbb{C}$. Let $K$ be an arbitrary field with $|K|>n$ (where $n \in \mathbb{N}$ is our matrix dimension) and suppose $T(I)=CD$ for some matrices $C$ and $D$. The claim is that there exists some invertible matrix $B$ such that $T: A \mapsto CB^{-1}ABD$ or $T: A \mapsto CB^{-1}A^tBD$.

  • We can begin by proving that $T$ is injective (and hence bijective) by the same argument as given by Wofsey. We change from $T$ to $T:=T'(.)=C^{-1}T(.)D^{-1}$ such that $T(I)=I$.

  • To show that $T$ preserves the rank of matrices, recall that a rank $m$ matrix can be written as $A=Q\Lambda$ where $\Lambda$ is nonsingular and $Q$ has precisely $m$ nonzero linear independent columns (the other columns being zero). With this in mind, we can see that for any matrix $A$ \begin{equation} \DeclareMathOperator\rank{rank}\rank(A)=\max \left\{ m \in \mathbb{N}\mid D \in M_n:\det (\lambda D + A)= c_{n-m}\lambda^{n-m}+\dotsb+c_n \lambda^n \text{ with }c_{n-m} \neq 0 \right\}. \end{equation} (If I'm correct, the assumption that $|K|>n$ is important here, in the sense that for a polynomial $P$ of degree $\leq n$ we have $P(\lambda)=0$ for all $\lambda \in K$ only if $P=0$. Hence polynomials $P$ are uniquely fixed by their evaluations $P(\lambda)$.)

But from the bijectivity of $T$ it follows that \begin{equation} \rank(A)=\max \left\{ m \in \mathbb{N}\mid D \in M_n:\det (\lambda D + A)=\det (T(\lambda D + A))=\det (\lambda T(D) + T(A))= c_{n-m}\lambda^{n-m}+\dotsb+c_n \lambda^n \text{ with }c_{n-m} \neq 0 \right\}= \max \left\{ m \in \mathbb{N}\mid D \in M_n:\det (\lambda D +T(A))= c_{n-m}\lambda^{n-m}+\dotsb+c_n \lambda^n \text{ with }c_{n-m} \neq 0 \right\}=\rank(T(A)). \end{equation}

  • The last modification to make is to revise the proof that $T$ maps projectors on projectors of equal rank. It is sufficient to prove that $A$ is a projector of rank $m$ iff $A$ has rank $m$ and $I-A$ has rank $n-m$.

Proving the rightward implication is easy. For the leftward implication: for any matrix $A$, we have the inclusion $\ker (A) \subset \DeclareMathOperator\Ran{Ran}\Ran (I-A)$ (easy exercise), which implies $\dim(\ker (A)) \leq \dim(\Ran (I-A))$ where equality is attained iff $\ker (A) = \Ran (I-A)$. By the dimension theorem, we then have that $n=\dim(\Ran(A))+\dim(\ker(A))\leq \dim(\Ran(A)) + \dim(\Ran(I-A))$ with equality iff $\ker (A) = \Ran (I-A)$. We conclude that \begin{equation} \rank(A)+\rank(I-A)=\dim(\Ran(A))+\dim(\Ran(I-A))=n \Rightarrow \ker (A) = \Ran (I-A) \Rightarrow A(I-A)=0 \Rightarrow A\text{ is a projector.} \end{equation}

$\endgroup$
5
  • $\begingroup$ For infinite fields, my argument can be adapted more simply as follows. First, my argument applies verbatim to algebraically closed fields. Now note that the identity $\det T(A)=\det(A)$ must also hold over the algebraic closure $\overline{K}$ of $K$ as well, since $M_n(K)$ is Zariski-dense in $M_n(\overline{K})$. Applying my argument in $\overline{K}$, it is easy to see that the $U$ and $V$ produced actually have entries in $K$. (In fact, because $\det(A)$ and $\det T(A)$ are polynomials of degree $\leq n$, you only need $|K|>n$.) $\endgroup$ Oct 18, 2015 at 22:10
  • $\begingroup$ Okay, but invoking the Jordan normal form (or whatever related thing) for this purpose feels a bit expensive. You seem to use it in a place where just the dimension theorem suffices. By the way, I suspect that the $|K|>n$-assumption is in fact also unnecessary. $\endgroup$
    – 5th decile
    Oct 18, 2015 at 22:16
  • $\begingroup$ Sure; your arguments are definitely nice and more elementary than mine. I guess the point of my previous comment is just that it is quite easy to reduce to the algebraically closed case by formal considerations (but your answer is still useful as providing a more elementary argument even in that case). $\endgroup$ Oct 18, 2015 at 22:29
  • $\begingroup$ Do we know if this is true without the $|K|>n$ hypothesis? $\endgroup$ Apr 26, 2020 at 13:46
  • $\begingroup$ I just read your good post (+1). Your characterization of $Rank(A)$ (line 14) seems -to me- valid over $F_2$ (for any $n\times n$ matrix)) -anyway, I have no counter-example-. Where do you use your hypothesis $|K|>n$ ? $\endgroup$
    – loup blanc
    Apr 29, 2020 at 13:04
3
$\begingroup$

Here is a more geometric proof. Put $V=\mathbb{C}^n$. Consider the embedding $i:\mathbb{P}(V)\times \mathbb{P}(V^*)\hookrightarrow \mathbb{P}(\mathrm{End}(V))$ which associates to $(x,x^*)$ the endomorphism $z\mapsto \langle x^*,z\rangle x\ $; its image is the locus of rank 1 endomorphisms. Since $T$ preserves the rank, it induces an automorphism of $\mathbb{P}(V)\times \mathbb{P}(V^*)$. Now $i$ is the embedding defined by the global sections of the line bundle $\mathcal{O}(1)\boxtimes \mathcal{O}(1)$; any automorphism of $\mathbb{P}(V)\times \mathbb{P}(V^*)$ preserves this line bundle, hence is induced by a unique automorphism of $\mathbb{P}(\mathrm{End}(V))$. These automorphisms are of the form $(x,x^*)\mapsto (u(x),{}^{t}\!v(x^*))$, with $u,v\in \mathrm{Aut}(V)$, or $(x,x^*)\mapsto ({}^{t}\!v(x^*),u(x))$, where $u$ (resp. $v$) is an isomorphism of $V$ onto $V^*$ (resp. $V^*$ onto $V$). It follows that $T(f)=ufv\ $ or $\ u{}^{t}\!fv$.

$\endgroup$
5
  • $\begingroup$ What can be said about trace preserving maps? $\endgroup$ Jun 9, 2019 at 14:04
  • $\begingroup$ Then you are just asking for linear endomorphisms of a vector space preserving a given linear form. This is an easy and rather uninteresting exercise. $\endgroup$
    – abx
    Jun 9, 2019 at 14:37
  • 1
    $\begingroup$ yes you are right. But what about if we give a bundle flavour to our question: we can consider a fiber bundle over the unit sphere of $V^*$. Each fober over a form $ \alpha \in V^*$ is all element of$ Gl(V)$ preserving $\alpha$ $\endgroup$ Jun 9, 2019 at 16:18
  • 1
    $\begingroup$ one can consider the similar construction to ontain a vector bundle and corresponding principal bundle. $\endgroup$ Jun 9, 2019 at 16:20
  • $\begingroup$ the same construction on $PV^*$ seems more interesting.what do you think? $\endgroup$ Jun 9, 2019 at 16:52
1
$\begingroup$

It seems (at least for me) that the condition $|K|>n$ —obtained by Vergilius— is useful only to show his characterization of the rank. We show that this result is valid for any field.

Proposition. Let $K$ be any field, $A\in M_n(K)$ and let $f:D\in M_n(K)\mapsto \DeclareMathOperator\degree{degree}\degree(\det(D+sA),s)$.

Then $\max_D f(D)=\DeclareMathOperator\rank{rank}\rank(A)$.

Proof. We may assume that $A=\DeclareMathOperator\diag{diag}\diag(N,U)$ where $N\in M_p(K)$ is nilpotent and in Jordan form and $U\in \operatorname{GL}_{n-p}(K)$.

i) $f(D)\leq \rank(A)$. Let $u=\rank(N)$; note that $\rank(A)=n-p+u$.

Consider one of the $n!$ elements of $f(D)$; with the $n-p$ last columns, we obtain at most $n-p$ times the factor $s$; with the $p$ first columns we obtain at most $u$ times $s$ (it's exactly the number of $s$ amongst the entries of $N$).

ii) There is $D$ s.t. $f(D)=\rank(A)$. We choose $D$ in the form $D=\diag(P,0)$. Then $f(D)=\det(P+sN)\det(sU)$. It remains to show that there is $P=[p_{i,j}]$ s.t. $\degree(P+sN)=u$.

Let $N=\diag(J_{i_1},\dotsc,J_{i_q})$ where $J_k$ is the nilpotent Jordan block of dimension $k$. Then it suffices to choose the $p_{i,j}$ equal to $1$ for the entries which are located at the bottom left of each block of Jordan and, otherwise, $0$.

Example: for $p=2+3$, $P+sN=\diag(\begin{pmatrix}0&s\\1&0\end{pmatrix},\begin{pmatrix}0&s&0\\0&0&s\\1&0&0\end{pmatrix})$.

$\endgroup$
3
  • $\begingroup$ Can we simplify this proof by using the fact that every $A \in M_n(K)$ can be written in the form $A = XSY$, where $X$ and $Y$ are invertible and $S = \mathrm{diag}(1,...,1,0,...,0)$ (with $\mathrm{rank}(A)$ ones on the diagonal)? I think that works over any field ($X$ and $Y$ just come from doing row and column operations on $A$, respectively), and it lets you assume that $A = \mathrm{diag}(1,...,1,0,...,0)$ instead of the more complicated block matrix form, right? $\endgroup$ May 9, 2020 at 2:30
  • $\begingroup$ Is the reference to a bound obtained/used by Thibaut Demaerel to the first answer by the user currently called Vergilius? $\endgroup$
    – LSpice
    Jan 20 at 18:59
  • 1
    $\begingroup$ @LSpice, yes it is. $\endgroup$
    – loup blanc
    Jan 20 at 23:06
0
$\begingroup$

Another proof

  1. Suppose $X$ is an $n$-dimensional $K$-vectorspace and $T:\DeclareMathOperator\End{End}\End(X) \to \End(X)$ preserves the determinant. Let $\{x_j\}_{1\leq j \leq n}$ be a basis of $X$ and $\{x_j^*\}_{1\leq j \leq n}$ its dual basis.

  2. $T$ is injective and preserves rank. For rank 1 matrices, one of the implications is that $$\forall j \in \{1,\ldots,n\}:T(x_1x_j^*) = y_1 \varphi_j^* \text{ and }T(x_jx_1^*)=y_j\varphi_1^*,\\ \text{or}, \forall j \in \{1,\ldots,n\}:T(x_1x_j^*) = y_j \varphi_1^*\text{ and }T(x_jx_1^*)=y_1\varphi_j^* \qquad(1)$$ where $\{y_j\}_{1\leq j \leq n}$ is a basis of $X$ ($\{y_j^*\}_j$ the corresponding dual basis) and $\{\varphi_j^*\}_{1\leq j \leq n}$ is a basis of $X^*$ ($\{\varphi_j\}_j$ the corresponding dual basis, keeping $X^{**}\simeq X$ in mind). We can restrict ourselves to the former scenario outlined in (1) by replacing $T$ in the latter scenario with $T\circ t$, where $t$ is a "coordinate-transpose", i.e. $t(x_jx_k^*):=x_kx_j^*$ and $t$ linear. If we then define the invertible matrices $U_1=\sum_{j=1}^n x_j y_j^*,\,V_1=\sum_{j=1}^n \varphi_jx_j^*$, then $T_1(.):=U_1T(.)V_1$ fixes both $\{x_1x_j^*\}_j$ and $\{x_jx_1^*\}_j$, i.e. $T_1$ fixes matrices whose non-zero entries occur only in the 1st row and/or first column. (Note that $T_1$ multiplies the determinant of its input by an uncertain factor $\det(U_1 V_1)\neq 0$, but later on the proof will indirectly show that this factor must be 1 anyway)

  3. $T_1$ maps rank 1 matrices to rank 1 matrices. At this stage, this implies that $\forall j,k \in \{2,\ldots,n\}:\,\exists a_{jk}\in \mathbb{C}:\,T_1(x_jx_k^*)=a_{jk}x_jx_k^*$ (test $T_1$ on $(x_1+x_j)x_k^*$ and on $x_j(x_1^*+x_k^*)$ to arrive at this conclusion), i.e. $T_1$ performs a 'pointwise' multiplication on the matrix entries of its input (viewed in the $\{x_j\}_j$-basis). To see that the $a$-coefficients all have to be equal to 1, test $T_1$ once more on the rank 1 matrix $(x_1+x_j)(x_1^*+x_k^*)$. So we conclude that $T_1$ is the identity map on $\End{X}$ and retracing our steps, we see that $T(.)\equiv U_1^{-1}(.)V_1^{-1}$ or $T(.)\equiv U_1^{-1}(t(.))V_1^{-1}$.

$\endgroup$
6
  • $\begingroup$ Could you describe how this relates to your other answer? $\endgroup$
    – LSpice
    Jan 20 at 18:55
  • 1
    $\begingroup$ @LSpice No relation, except maybe that I don't elaborate in this answer why $T$ preserves the rank while there is a proof of that in the other one (it's a basic lemma). $\endgroup$
    – 5th decile
    Jan 20 at 18:57
  • $\begingroup$ I guess I just meant to ask, does this supplant your prior answer, or is it an independent proof? $\endgroup$
    – LSpice
    Jan 20 at 18:59
  • 1
    $\begingroup$ @LSpice An independent proof $\endgroup$
    – 5th decile
    Jan 20 at 19:02
  • 1
    $\begingroup$ @LSpice Yes, that identification. My proof fixes a basis for $V$ once and for all. I'll try to take care of the double use of $V$ when I'm home. Also, I just noted that I have to treat the case $n=1, n+1=2$ a bit separately at the end of point 4. and in point 5. (the decision to take the transpose or not is taken differently for that $n$) $\endgroup$
    – 5th decile
    Jan 20 at 19:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.